For training NLP, Vision, and other deep learning models in the cloud, you must choose between On-Demand and Spot compute instances, and it’s a big choice.
Spell eliminates the need for costly on-demand compute instances for long-running model training jobs by transparently combining sequences of short-running, highly discounted “spot” instances into a single virtual on-demand instance that enables long-running training jobs without administrative overhead, at a fraction of the cost of dedicated instances.
With Spell’s virtual on-demand instances, when a spot instance is preempted, Spell auto-recovers the data and persists it to object storage, rather than in a costlier block storage volume. The next available instance is then automatically selected and started, and the training continues with no user intervention needed.
Spell’s unique compute infrastructure automation is enabling our customers to save 66% or more on cloud compute cost for deep learning. Read our blog article about how you can too. Or contact us to learn more.